Technology & RAG
How our AI works
Every AI-powered genRAG.AI card comes with AURA, a private AI assistant that answers visitors from the cardholder's own information. On genrag.ai, RAGA answers questions about our products the same way. This page explains how answers are made, how each card's data stays separate, and how we stop the AI from making things up.
1. AI Architecture
AURA uses retrieval-augmented generation (RAG): it first finds the right facts on the card, then writes the answer from those facts only.
- A knowledge index for each card. The card's details, services and prices, FAQs, knowledge base and PDF brochures are split into short passages and indexed. Search understands typos, synonyms and questions typed in English, Tamil, Tanglish, Hindi, Hinglish, Bengali, Kannada and Telugu.
- Instant answers first. Greetings, thanks and clear questions such as contact details, timings or a service list are answered straight from the card in milliseconds, without an AI model.
- AI for natural replies. For other questions, the best matching passages and a short draft are sent to an AI model, which writes a friendly reply in the visitor's language. We use free-tier models (Google Gemini first, then Groq, Cerebras or OpenRouter) and switch automatically if one is slow, busy or near its limit, so answers keep flowing.
- Voice. Visitors can speak their question and hear the reply. Speech is turned into text by the visitor's own browser; we receive only the text.
- Private mode. A cardholder can switch AURA to answer only from the built-in search engine, so no question leaves our server.
2. Tenant Isolation
Each cardholder's information is kept separate from everyone else's.
- One card, one index. A question to one card's AURA is answered only from that card's published details, FAQs, knowledge base and brochures. It can never search or quote another subscriber's data.
- Only what is needed leaves the server. When an AI model is used, it receives only that card's matching passages, the visitor's question and the last few messages of that conversation.
- No chat transcripts stored. We keep a count of questions per card, not the conversation text. Recent replies may be held in server memory for up to 12 hours so repeated questions are answered faster.
- Separated records. Profiles, uploads and enquiries are stored per account and per card; enquiries reach only that cardholder.
- Location limits. A cardholder can restrict their profile to up to 3 locations; visitors outside them can't open the card or its AI.
3. Zero-Hallucination Policy
AURA and RAGA must never invent facts. These rules are enforced in code, not just requested from the AI:
- Answers come only from the card. The AI is told to answer only from the card's own information and never from general knowledge or guesses.
- Every reply is fact-checked. Before a reply is shown, every number (prices, timings, phone numbers, years, quantities), email address and link in it must appear in the card's information or the conversation. If anything doesn't match, the AI reply is discarded and the verified answer from the search engine is shown instead.
- Honest when unsure. If the card doesn't have the answer, AURA says so and offers the cardholder's WhatsApp, phone or enquiry form, rather than guessing.
- Numbers copied exactly. Prices, timings and contact numbers are reproduced as written by the cardholder.
- Safe by design. Visitor messages are treated as questions, never as instructions, so they can't change these rules.
AI replies can still be imperfect. Please don't type passwords, payment details or health or other sensitive personal information into a chatbot. See our Privacy Policy for full details.